Dice Relevancy Feedback
Dice.com's solr plugins for performing personalized search, and recommendations (via the relevancy feedback plugin) and conceptual / semantic search (via the unsupervised feedback plugin).
Building the Plugin
A pre-built jar file can be found in the ./target
folder. The project contains a maven pom.xml file which can also be used to build it from source.
Supported Solr versions
- Solr 5.4 (see branch)
- Solr 6.3 (see branch) also master
If there is a particular version of Solr you need this for, please create a GitHub issue and I'll see what I can do. To manually compile it for a specific version, use maven to compile the plugins using the pom.xml file, and update the versions of the solr and lucene libraries in that file, and use maven to pull in those dependencies. Then fix any compilation errors.
Importing into SOLR
Please see the official SOLR guidelines for registering plugins with solr. This basically involves simply dropping the jar file into one of the folders that Solr checks for class and jar files on core reload.
Relevancy Feedback Plugin
An example request handler configuration for the solrconfig.xml is shown below, with comments outlining the main parameters:
<requestHandler name="/rf" class="org.dice.solrenhancements.relevancyfeedback.RelevancyFeedbackHandler">
<lst name="defaults">
<str name="omitHeader">true</str>
<str name="wt">json</str>
<str name="indent">true</str>
<!-- Regular query configuration - query parser used when parsing the rf query-->
<str name="defType">lucene</str>
<!-- fields returned -->
<str name="fl">jobTitle,skill,company</str>
<!-- fields to match on-->
<str name="rf.fl">skillFromSkill,extractTitles</str>
<!-- field weights. Note that term weights are normalized so that each field is weighted exactly in this ratio
as different fields can get different numbers of matching terms-->
<str name="rf.qf">skillFromSkill^3 extractTitles^4.5</str>
<int name="rows">10</int>
<!-- How many terms to extract per field (max) -->
<int name="rf.maxflqt">10</int>
<bool name="rf.boost">true</bool>
<!-- normalize the weights for terms in each field (custom to dice rf, not present in solr MLT) -->
<bool name="rf.normflboosts">true</bool>
<!-- Take the raw term frequencies (false) or log of the term frequenies (true) -->
<bool name="rf.logtf">true</bool>
<!-- Minimum should match settings for the rf query - determines what proportion of the terms have to match -->
<!-- See Solr edismax mm parameter for specifics -->
<bool name="rf.mm">25%</bool>
<!-- Returns the top k terms (see regular solr MLT handler) -->
<str name="rf.interestingTerms">details</str>
<!-- Turns the rf query into a boost query using a multiplicative boost, allowing for boosting -->
<str name="rf.boostfn"></str>
<!-- q parameter - If you want to execute one query, and use the rf query to boost the results (e.g. for personalizing search),
pass the user's query into this parameter. Can take regular query syntax e.g.rf.q={!edismax df=title qf=.... v=$qq}&qq=Java
The regular q parameter is reserved for the rf query (see abpve)
-->
<str name="q"></str>
<!-- Query parser to use for the q query, if passed -->
<str name="defType"></str>
<!-- rf.q parameter - Note that the regular q parameter is used only for personalized search scenarios, where you have a main query
and you want to use the rf query generated to boost the main queries documents. Typically the rf.q query is a query that identifies
one or more documents, e.g. rf.q=(id:686867 id:98928980 id:999923). But it cam be any query. Note that it will process ALL documents
matched by the q query (as it's intended mainly for looking up docs by id), so use cautiously.
-->
<str name="rf.q"></str>
<!-- query parser to use for the rf.q query -->
<str name="rf.defType"></str>
<!-- Settings for personalized search - use the regular parameter names for the query parser defined by defType parameter -->
<str name="df">title</str>
<str name="qf"> company_text^0.01 title^12 skill^4 description^0.3</str>
<str name="pf2">company_text^0.01 title^12 skill^4 description^0.6</str>
<!-- Content based recommendations settings (post a document to the endpoint in a POST request). The stream.body and stream.head are form parameters
You can send them in a GET request, but a POST handles larger data. If you have really large documents, you will need to change the buffer settings
so that the request doesn't blow the buffer limits in Solr or your web server.
-->
<!-- Fields used for processing documents posted to the stream.body and stream.head parameters in a POST call -->
<str name="stream.head.fl">title,title_syn</str>
<str name="stream.body.fl">extractSkills,extractTitles</str>
<!-- pass a url in this parameter for Solr to download the webpage, and process the Html using the fields configured in the stream.qf parameters -->
<str name="stream.url"></str>
<!-- Note that we have two different content stream fields to pass over in the POST request. This allows different analyzers to be appkied to each.
For instance, we pass the job title into the stream.head field and parse out job titles, while we pass the job description to the stream.head parameter
to parse out skills -->
<!-- Pass the document body in this parameter as a form parameter. Analysed using the stream.body.fl fields-->
<str name="stream.body"></str>
<!-- Pass the second document field in this parameter. Analysed using the stream.head.fl fields-->
<str name="stream.head"></str>
<!-- Specifies a separate set of field weights to apply when procesing a document posted to the request handler via the
stream.body and stream.head parameters -->
<str name="stream.qf">extractSkills^4.5 extractTitles^2.25 title^3.0 title_syn^3.0</str>
</lst>
</requestHandler>
Example Request
Example Response
{
"match":{
"numFound":2,
"start":0,
"docs":[
{
"id":"a2fd2f2e34667d61fadcdcabfd359cf4",
"title":"Console AAA Sports Video Game Programmer.",
"skills":["Sports Game Experience a plus.",
"2-10 years plus Console AAA Video Game Programming Experience"],
"geocode":"38.124447,-122.55051",
"city":"Novato",
"state":"CA"
},
{
"id":"11f407d319d6cc707437fad874a097c0",
"title":"Game Engineer - Creative and Flexible Work Environment!",
"skills":["3D Math",
"Unity3d",
"C#",
"3D Math - game programming",
"game programming",
"C++",
"Java"],
"geocode":"33.97331,-118.243614",
"city":"Los Angeles",
"state":"CA"
}
]
},
"response":{
"numFound":5333,
"start":0,
"docs":[
{
"title":"Software Design Engineer 3 (Game Developer)",
"skills":["C#",
"C++",
"Unity"],
"geocode":"47.683647,-122.12183",
"city":"Redmond",
"state":"WA"
},
{
"title":"Game Server Engineer - MMO Mobile Gaming Start-Up!",
"skills":["AWS",
"Node.JS",
"pubnub",
"Websockets",
"pubnub - Node.JS",
"Vagrant",
"Linux",
"Git",
"MongoDB",
"Jenkins",
"Docker"],
"geocode":"37.777115,-122.41733",
"city":"San Francisco",
"state":"CA"
},...
]
}
}
Unsupervised Feedback (Blind Feedback) Plugin
An example request handler configuration for the solrconfig.xml is shown below, with comments outlining the main parameters:
<requestHandler name="/ufselect" class="org.dice.solrenhancements.unsupervisedfeedback.UnsupervisedFeedbackHandler">
<lst name="defaults">
<str name="omitHeader">true</str>
<str name="wt">json</str>
<str name="indent">true</str>
<!-- Regular query configuration -->
<str name="defType">edismax</str>
<str name="df">title</str>
<str name="qf">title^1.5 skills^1.25 description^1.1</str>
<str name="pf2">title^3.0 skills^2.5 description^1.5</str>
<str name="mm">1</str>
<str name="q.op">OR</str>
<str name="fl">jobTitle,skills,company</str>
<int name="rows">30</int>
<!-- Unsupervised Feedback (Blind Feedback) query configuration-->
<str name="uf.fl">skillsFromskills,titleFromJobTitle</str>
<!-- How many docs to extract the top terms from -->
<str name="uf.maxdocs">50</str>
<!-- How many terms to extract per field (max) -->
<int name="uf.maxflqt">10</int>
<bool name="uf.boost">true</bool>
<!-- Relative per-field boosts on the extracted terms (similar to edismax qf parameter -->
<!-- NOTE: with uf.normflboosts=true, all terms are normalized so that the total importance of each
field on the query is the same, then these relative boosts are applied per field-->
<str name="uf.qf">skillsFromskills^4.5 titleFromJobTitle^6.0</str>
<!-- Returns the top k terms (see regular solr MLT handler) -->
<str name="uf.interestingTerms">details</str>
<!-- unit-length norm all term boosts within a field (recommended) - see talk for details -->
<bool name="uf.normflboosts">true</bool>
<!-- use raw term clounts or log term counts? -->
<bool name="uf.logtf">false</bool>
</lst>
</requestHandler>
Example Request
Example Response
{
"match":
{
"numFound":7729,
"start":0,
"docs":[
{
"title":"NLP/Machine Learning Engineer",
"skills":["Linux",
"NLP (Natural Language Processing)",
"SQL",
"Bash",
"Python",
"ML (Machine Learning)",
"JavaScript",
"Java"],
"geocode":"42.35819,-71.050674",
"city":"Boston",
"state":"MA"
},
{
"title":"Machine Learning Engineer",
"skills":["machine learning",
"java",
"scala"],
"geocode":"47.60473,-122.32594",
"city":"Seattle",
"state":"WA"
},
{
"title":"Machine Learning Engineer - REMOTE!",
"skills":["Neo4j",
"Hadoop",
"gensim",
"gensim - C++",
"Java",
"R",
"MongoDB",
"elastic search",
"sci-kit learn",
"Python",
"C++"],
"geocode":"37.777115,-122.41733",
"city":"San Francisco",
"state":"CA"
},...
]
}
Isn't this just the MLT Handler?
While it is loosely based on the Solr MLT handler code and algorithm (which is just the Rocchio algorithm), there are some key differences in the algorithm design. The MLT handler takes the top k terms across all configured fields when constructing the MLT query. If you have a field that has a broader vocabulary than the other fields, the average document frequency of a term will be lower than in other fields with smaller vocabularies. This means that these terms will have high relative idf scores and tend to dominate the top terms selected by the Solr MLT handler. Our request handler takes the top k terms per field. It also ensure that that no matter how many terms are matched per field (up to the configured limit), that field has the same weighting in the resulting query as all other fields, before the field specific weights specified in the rf.qf parameter are applied. This is the second problem with the Solr MLT handler that we address. We also provide a lot of extra functionality. We allow for passing in of content streams, matching against multiple documents (more like 'THESE' as opposed to more like 'this'), applying the boost query parser to the resulting MLT query to allow for any arbitrary solr boost to be applied (multiplicative). And we support the mm parameter, so we can force documents to come back that only match a set % of the top terms.
Important Considerations When using for Personalized Search
If you wish to use this to perform search personalization, as demonstrated in my Lucene Revolution 2017 talk, you need to pass in the user's current search query using the regular q parameter, and the information used to generate the rocchio query is passed via the rf.q parameter (when using documents to generate the Rocchio query) or via the content stream parameters (rf.stream.head and rf.stream.body, which take strings of content). Note however, that the boosts applied to the terms in the rocchio query are not of comparative weights to those in your user query, due to the process of normalization that the algorithm applies. So you will need to experiment with different rf.qf values until you find the right level of influence on your query, based on your search configuration. Also, given that the rocchio query generated for each user is likely the same across the user's search session (depending on your use case of course), a more efficent way of using this to do personalization is simply to use the RF handler to generate the rochio query for you once when the user logs in, cache this query, and then use it as a boost query (within your regular search request handler) for personalizing subsequent user searches. The handler returns the rocchio query in the rf.query parameter in the response. If you want to use the handler just to get the query (and not execute the search), you can set the rows parameter to 0. You can also iterate over the set of 'interesting terms' returned by the algorithm, along with their weights, if you set rf.interestingTerms=details, and use this to build your boost query.
Potential Enhancements
Aside from ensuring this works with more versions of solr (please leave feedback as to which versions you all want), there are a number of possible enhancements:
- Relevancy Feedback Handler Allow the learning of negative terms from the negative examples (if supplied - needs a separate query parameter), then implement using negative boosting. Another enhancement would be to allow the max terms per field (rf.maxflqt) to be specified on a per field basis, so that you can vary the max number of terms extracted by field.
- Unsupervised Feedback (Blind Feedback) Use the positional relevance model detailed in this paper: http://dl.acm.org/citation.cfm?id=1835546. This uses only terms found near the query's terms in the document, as these are generally more relevant than using the whole document. The highlighter component can presumably be used as a reference to determine how to get this information from the postings list, or maybe even used directly to get this information.
Contact Details
If you have a feature request, please submit it to the issues list. If you have questions, that is also a good place to post them, but you can also reach out to me at simon.hughes@dice.com if you don't here back.